README.md

Geometric Shapes Dataset

This dataset contains procedurally generated images of various geometric shapes with corresponding captions. It's designed for educational purposes and testing of diffusion models.

Dataset Overview

• Content: 100,000 images of geometric shapes with detailed metadata
• Image size: 512x512 pixels
• Format: PNG images with CSV metadata
• Features: Various shapes, colors, sizes, and descriptive captions
• Purpose: Educational use for training and testing single-step diffusion models

Dataset Creation

The dataset was generated using the included Python script (create_dataset.py), which was created with the assistance of Claude AI (Anthropic). The script generates a wide variety of geometric shapes with different properties:

• Shapes: 50+ different shapes including basic shapes (circle, square, triangle) and more complex ones (hypocycloid, lemniscate, star)
• Colors: 30 different colors with their hex values
• Sizes: Randomly scaled from 10% to 100% with corresponding descriptive terms
• Captions: Diverse, randomly generated descriptions of the shapes in various formats

Usage

Using the Dataset

This dataset is designed for educational purposes only. It's particularly useful for:

• Learning to build diffusion models from scratch
• Experimenting with simple image-to-text and text-to-image models
• Testing model performance on synthetic data
• Educational demonstrations of computer vision concepts

OPTIONAL - If you want to createw youe own images

If you want to generate your own version of the dataset or modify the parameters:

1. Install the required dependencies:

   pip install numpy matplotlib tqdm
   

2. Run the script: create a folder "geometric_shapes", and save create_dataset.py the in the folder, and execute it

   python create_dataset.py
   

   Generating Images: 100%|█████████████████████████████████████████████████| 100000/100000 [07:36<00:00, 218.96it/s]
   Generated 100000 images in 456.89 seconds
   Size: 100000 images
   

3. By default, the script will:

   • Generate 100,000 images in the dataset/images directory
   • Create a dataset/metadata.csv file with image filenames and captions
   • Use 10 CPU cores for parallel processing

4. You can modify these parameters in the __main__ section of the script:

   generate_images_multiprocess(
       num_images=100000,  # Change this number to generate fewer/more images
       max_workers=10,     # Change based on your CPU cores
       img_size=(512, 512) # Modify image dimensions
   )
   

5. Ensure your dataset is organized as follows:

geometric_shapes/
├── README.md
├── create_dataset.py
├── dataset/
│   ├── images_001/
│   │   ├── image1.jpg
│   │   ├── image2.jpg
│   │   ├── ...

6. Upload Instructions

   Log in to Hugging Face:

   huggingface-cli login
   

   Upload the dataset:

   cd .. && huggingface-cli upload-large-folder anokimchen/geometric_shapes geometric_shapes  --repo-type dataset && cd geometric_shapes
   

License

This dataset is provided for educational and research purposes only.

Acknowledgements

• Code generated with the assistance of Claude (Anthropic)
• Dataset created for learning purposes to build and test single-step diffusion models

Citation

If you use this dataset in your research or educational projects, please include a reference to this repository.